---
title: Realtime, agents, and streaming
description: Operate generations, limits, Redis state, and strict client deployments.
pagination: false
sidebar:
  icon: radio
---

Keep health checks and controls responsive while clients observe long-running jobs, agents, or streams. Observation
does not consume finite function concurrency. A run remains bound to its original generation;
another generation may serve compatible journal reads and persist controls for the owner, but cannot
migrate execution or continuation. Forced retirement reports `GENERATION_UNAVAILABLE` without replay.

> **Before you begin**
>
> Apply this operational guidance after [realtime provider
> installation](/docs/modules/realtime/providers), the selected [native agent
> setup](/guides/ai/build-an-offline-agent), and [frontend installation](/docs/modules/client).
> Selecting a channel does not select a model engine, and frontend installation does not install
> backend providers.

Deploy frontend and backend public-contract changes together. Even an additive public change blocks
an older browser under exact fingerprint matching. Show an application-updated state, preserve
unknown-operation metadata, and do not auto-reload while a mutation or control is unresolved.

Tune aggregate connection, partition, retained-byte, run, thread, and control limits before public
exposure. Per-principal overload is 429; application/provider overload is 503. Redis-backed guarantees
depend on persistence configuration, and provider resets must advance the provider epoch.